A global-local feature collaborative enhancement spatial domain small target detection method

Abstract

A method for detecting small targets in the airspace with global-local feature co-enhancement is presented, belonging to the field of airspace small target detection technology. This method strengthens the collaborative representation capability of global and local features by constructing a Contextual Feature Aggregation Module (CFFM), a Tiny Target Region Focusing Module (TTFM), and a Dynamic Feature Aggregation Module (DFAM). The network embeds CFFM in the backbone to integrate channel and spatial context information, introduces an additional branch and adds TTFM to enhance local perception of small targets, and sets DFAM to dynamically fuse multiple receptive field features before the detection head. This invention is validated using a self-built ASODD dataset and the publicly available VisDrone2019-DET dataset. Compared with existing mainstream algorithms, the mAPsmall index for small target detection is significantly improved, solving the problems of low accuracy and high false positive / false negative rates in airspace small target detection. It is suitable for efficient detection of small targets such as drones, birds, and kites in complex airspace environments.

Description

Technical Field

[0001] This invention belongs to the field of small airspace target detection technology, specifically involving a deep learning-based global-local feature collaborative enhancement detection method, which is applicable to scenarios such as anti-drone intrusion and airspace security monitoring, and can achieve efficient and accurate detection of small aerial targets such as drones, birds, and kites. Background Technology

[0002] With the widespread use of civilian and commercial drones, airspace security issues are becoming increasingly prominent, especially in sensitive areas such as airports and military bases, which face the risk of drone intrusion. Small target detection in the airspace is a core component of counter-drone systems, requiring early detection and handling of small targets at long distances against complex backgrounds. However, key targets in the airspace (such as drones and birds) are typically small in size, occupy few image pixels, and lack sufficient texture and shape information, posing a significant challenge to traditional target detection algorithms.

[0003] Existing target detection technologies are divided into two-stage detectors and one-stage detectors. Two-stage detectors have high accuracy but slow detection speed, making it difficult to meet the needs of real-time monitoring; one-stage detectors have high detection efficiency, but when dealing with small targets in the spatial domain, they suffer from low detection accuracy and high false alarm and false negative rates because they cannot simultaneously capture global context information and extract local detailed features.

[0004] Existing improvement methods either fuse multi-scale features through feature pyramids or enhance local perception through attention mechanisms. However, these methods generally suffer from insufficient global-local feature synergy: the sparse local features of small targets lack global contextual constraints and are easily confused with background interference information, resulting in limited detection performance. Therefore, designing a detection network that can enhance the synergistic representation of global and local features is of great significance for improving the detection performance of small targets in the spatial domain. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method for detecting small targets in the airspace with global-local feature co-enhancement. Validated using a self-built ASODD dataset and the publicly available VisDrone2019-DET dataset, the method significantly improves the mAPsmall index for small target detection compared to existing mainstream algorithms. It solves the problems of low accuracy and high false positive / false negative rates in small target detection in the airspace and is suitable for efficient detection of small targets such as drones, birds, and kites in complex airspace environments.

[0006] The technical solution adopted by this invention to solve its technical problem is: A spatial small target detection method with global-local feature co-enhancement includes the following steps: S1. Constructing the overall network architecture: The backbone network processes the input image through deep feature extraction and generates a high-level feature representation; the neck uses a PAN bidirectional feature pyramid network for feature fusion; the detection head is set with four detection branches, corresponding to the detection of large targets, medium targets, small targets and tiny targets respectively; S2. Embed the Context Feature Aggregation Module (CFFM) in the backbone network: extract features in parallel through multi-scale convolution, and combine global average pooling (GAP) and global max pooling (GMP) to integrate channel and spatial context information to construct a global feature-dense representation. S3. An additional micro-target branch is introduced between the backbone network and the neck, and a micro-target region focusing module TTFM is embedded: local context information and detailed features are extracted in collaboration between multi-head self-attention branches and convolutional branches to enhance the accuracy of micro-target localization. S4. Add a dynamic feature aggregation module DFAM in front of the detection head: adopt a multi-branch heterogeneous convolution design, dynamically weighted and fused local details and global context features of different receptive fields, and optimize feature representation in dense target scenes; S5. Constructing a training dataset: Combining the publicly available VisDrone2019-DET dataset with the self-built ASODD airspace small target dataset, the ASODD dataset contains 8241 labeled images of three types of targets: birds, kites, and drones. S6. Model Training and Detection: The model is trained using the stochastic gradient descent (SGD) optimizer. After feature processing by each module, the detection head outputs the target category, bounding box, and confidence score to complete the detection of small targets in the spatial domain.

[0007] Furthermore, in S2, the CFFM includes a multi-scale parallel convolution branch and an attention enhancement branch: the multi-scale parallel convolution branch extracts multi-scale features, namely, it uses a 3×3 standard convolution, a 5×5 standard convolution, and a large receptive field convolution branch composed of a cascaded 1×7 depthwise separable convolution and a 7×1 depthwise separable convolution, which process the input features in parallel; at the same time, a direct cross-layer connection from the input to the multi-scale fused features is designed to share information between different layers; at the end of the multi-scale parallel convolution branch, a 1×1 convolution is used to perform channel fusion and dimension unification on the parallel extracted multi-scale features; in the attention enhancement branch, global channel features are first extracted using both global average pooling (GAP) and global max pooling (GMP), and then the pooling results are summed and normalized by Softmax to generate channel attention weights; this branch contains two sub-paths: the first sub-path performs matrix multiplication of the channel attention weights and the multi-scale enhanced features to generate global spatial modulation features; the second sub-path first reduces the dimensionality of the channel attention weights using a 1×1 convolution, and then uses a Sigmoid function. The spatial attention weights are activated and multiplied element-wise with the original features (Hadamard product), followed by a 1×1 convolution to transform the spatially modulated features. Finally, the outputs of the two sub-paths are summed and then residually concatenated with the original input features of the module to form the final output of the module.

[0008] Furthermore, in S3, within the TTFM, the input features first undergo dimensionality transformation through feature reshaping and a fully connected layer. Then, a parallel branch employs a shifted window multi-head self-attention module (SW-MSA) to capture local contextual information. SW-MSA models local long-range dependencies through self-attention mechanisms within the window and expands the effective receptive field using a window shifting strategy, enhancing the model's ability to locate small targets and providing a clear local view for small target detection. Afterward, a multilayer perceptron (MLP) completes feature transformation and nonlinear mapping. In another parallel branch, the convolutional branch, convolutional operations further enhance local detail feature extraction: first, a 1×1 convolution is used for channel-dimensional information fusion and compression; then, a 5×5 convolution expands the local receptive field, capturing a wider range of detail information, especially helpful in accurately locating the boundaries of small targets in complex backgrounds. The output features of the two branches are added and fused. Finally, the fused features undergo dimensionality adjustment through feature reshaping and a 1×1 convolution to adapt to the processing requirements of subsequent network layers, forming the module's final output.

[0009] Furthermore, in S4, within the DFAM, the input feature map is divided into two main paths. One main path is the basic feature branch: the input feature map first undergoes a 1×1 convolution to complete channel transformation, and then is multiplied by a weight 1-α as a global feature contribution. The other main path is the core of multi-scale feature extraction and fusion, with the following process: First, after dimensionality reduction via a 1×1 convolution, the input feature map is divided into two sub-paths. One sub-path directly uses a 1×1 convolution, and the other sub-path uses a 5×1 convolution with a stride of 2 to extract vertical features. The outputs of the two sub-paths are multiplied element-wise, and then divided into two sub-paths again. One sub-path uses a 1×1 convolution, and the other sub-path uses a 1×5 convolution with a stride of 2 to extract horizontal features. The outputs of these two sub-paths are multiplied element-wise again. Subsequently, this result is concatenated with the outputs of the other two sub-paths, which are respectively the convolution branch after the element-wise multiplication of the first part and the 5×5 convolution branch with a stride of 2. The concatenated features are then processed by a 4×C2 channel 1×1 convolution. The convolutional compression and fusion are then multiplied by a weight α. Finally, the feature maps of the two main paths are added element by element to obtain the output. α is adaptively updated during training, and the final output is a comprehensive feature that takes into account both local details and global context.

[0010] Furthermore, in S5, the ASODD dataset is constructed as follows: N1 bird and kite samples are selected from the COCO dataset, N2 drone samples are extracted from the drone-detection-rchy7_dataset, and N3 drone samples are collected autonomously, for a total of N1+N2+N3 images. The target categories and bounding boxes are uniformly labeled, and the images are divided into training, validation, and test sets according to a set ratio. All images are preprocessed: resized to 640×640 pixels, and data augmentation techniques such as random flipping, brightness adjustment, and Gaussian blur are used to improve the model's generalization ability.

[0011] In S6, the training parameters are set as follows: PyTorch 1.18 framework, CUDA 12.4, dual NVIDIA GeForce RTX 3080 24GB graphics cards are used; the initial learning rate is set, the cosine annealing learning rate scheduling strategy is adopted, and the momentum, weight decay, batch size, and maximum training epoch are set. The cross-entropy loss function and GIoU loss function are used for joint optimization. The detection process is as follows: the image to be detected is input into the trained GLFE-TDN network. After the backbone network feature extraction, CFFM global enhancement, TTFM local enhancement, and DFAM feature fusion, the target detection results are output by the four detection heads. Non-maximum suppression (NMS) is used to remove redundant detection boxes. Finally, the target category, bounding box, and confidence score with a confidence score greater than the threshold are output.

[0012] The beneficial effects of this invention are mainly reflected in the following aspects: It innovatively proposes the GLFE-TDN network architecture, which, through the precise collaboration of the Contextual Feature Aggregation Module (CFFM), the Tiny Target Region Focusing Module (TTFM), and the Dynamic Feature Aggregation Module (DFAM), constructs a full-process feature optimization link of "global perception - local enhancement - dynamic fusion," fundamentally solving the core pain point of insufficient global-local feature collaboration in small target detection in the airspace. CFFM deeply integrates channel and spatial context information using multi-scale convolution and global pooling techniques, laying a solid global semantic foundation. TTFM focuses on the local details and contextual associations of tiny targets through multi-head self-attention and convolutional dual-branch collaboration, significantly improving the positioning accuracy in complex backgrounds. DFAM uses a multi-branch heterogeneous convolution design to dynamically weight and fuse different receptive field features, achieving efficient collaborative expression of local details and global context. The three modules are progressively layered and each performs its own function, completely breaking through the technical bottleneck of existing algorithms that struggle to simultaneously capture global information and extract local details. A high-precision small target detection model in the airspace is constructed, achieving efficient identification and positioning of small targets such as drones, birds, and kites. To further enhance the model's generalization ability and practical value, a multi-source data fusion training mechanism was designed. Through joint training with public datasets and the self-built ASODD dataset, combined with data augmentation techniques such as random flipping and brightness adjustment, the model's stable detection performance in different spatial scenarios is ensured. Attached Figure Description

[0013] Figure 1 This is a flowchart of the spatial small target detection method with global-local feature co-enhancement according to the present invention; Figure 2 This is a schematic diagram of the overall structure of the GLFE-TDN model of the present invention; Figure 3 This is a schematic diagram of the structure of the CFFM of the present invention; Figure 4 This is a schematic diagram of the structure of the TTFM of the present invention; Figure 5 This is a schematic diagram of the structure of the TTFM of the present invention. Detailed Implementation

[0014] The present invention will now be further described with reference to the accompanying drawings.

[0015] Reference Figure 1 and Figure 2 A spatial small target detection method with global-local feature co-enhancement includes a backbone network, a feature fusion neck, a detection head, and three core functional modules. The detection method includes the following steps: S1. Constructing the overall network architecture: The backbone network adopts a deep convolutional neural network, which is composed of alternating stacks of A2C2F and C3K2 modules to achieve deep feature extraction of the input image and generate high-level feature representations; the neck adopts a PAN bidirectional feature pyramid network, which fuses feature maps of different levels through upsampling, convolution, and concat operations to enhance the multi-scale feature expression capability; the detection head is set with four detection branches, corresponding to the detection of large, medium, small, and tiny targets, respectively. Each branch contains a Detect module, which is responsible for outputting the target's class probability, bounding box coordinates, and confidence score.

[0016] S2. Implementation of the Contextual Feature Aggregation Module (CFFM): CFFM is embedded between the A2C2F modules of the backbone network. The left-hand multi-scale feature extraction unit uses 3×3 convolution, 5×5 convolution, and cascaded 1×7 and 7×1 separable convolutions to process input features in parallel, while setting direct cross-layer connections to ensure information sharing between low-level and high-level features. The right-hand global feature integration unit performs GAP and GMP operations on the input features to extract global statistical information in the channel dimension. After Softmax normalization, the first branch combines the global channel features with the original features through matrix multiplication to generate global spatial features. The second branch fuses the global channel features with the spatial features through 1×1 convolution and then through Hadamard product operation. Finally, the fused features are added to the original input features to output global enhanced features, effectively improving the model's global semantic perception ability for small targets in complex backgrounds. The CFFM structure design in this embodiment is shown in Figure 3, including a multi-scale parallel convolution branch and an attention enhancement branch. The multi-scale parallel convolution branch extracts multi-scale features, employing 3×3 standard convolutions, 5×5 standard convolutions, and a large receptive field convolution branch composed of cascaded 1×7 and 7×1 depthwise separable convolutions, all processing input features in parallel. A direct cross-layer connection from the input to the multi-scale fused features is also designed to share information across different layers. At the end of the multi-scale parallel convolution branch, a 1×1 convolution is used to perform channel fusion and dimensionality unification on the parallel-extracted multi-scale features. In the attention enhancement branch, global channel features are first extracted using both Global Average Pooling (GAP) and Global Max Pooling (GMP). The pooling results are then summed and normalized using Softmax to generate channel attention weights. This branch contains two sub-paths: the first sub-path performs matrix multiplication between the channel attention weights and the multi-scale enhanced features to generate global spatial modulation features; the second sub-path first reduces the dimensionality of the channel attention weights using a 1×1 convolution, then applies a Sigmoid function. The spatial attention weights are activated and multiplied element-wise with the original features (Hadamard product), followed by a 1×1 convolution to transform the spatially modulated features. Finally, the outputs of the two sub-paths are summed and then residually concatenated with the original input features of the module to form the final output of the module.

[0017] The CFFM described above has significant adaptability advantages for small target detection in the airspace. It is suitable for small targets in the airspace with a small proportion of distant pixels, complex backgrounds, and easy fusion between targets and backgrounds. Its multi-scale convolutional parallel structure can accurately capture the basic features of small targets in the airspace at different scales, and is suitable for the size differences of different targets in the airspace such as drones, birds, and kites. Cross-layer connections can effectively preserve key features such as low-level edges and contours of small targets in the airspace, avoiding the loss of target details in deep feature extraction. The global pooling and dual-branch fusion methods can deeply integrate channel and spatial context information to construct a dense representation of global features. This can effectively suppress interference information from complex backgrounds in the airspace, enhance the model's global semantic perception ability of differentiated small targets in the airspace, and solve the problem that small targets in the airspace are easily confused by the background due to the lack of global context constraints.

[0018] S3. Implementation of the Tiny Target Region Focusing Module (TTFM): An additional tiny target branch is introduced between the backbone network and the neck region. This branch retains rich low-level semantic information and embeds TTFM for local feature enhancement. The multi-head self-attention branch of TTFM adopts the SW-MSA module with a window size of 7×7. The receptive field is expanded through a window shifting strategy to model long-range dependencies within the local region. The convolutional branch uses 1×1 convolution and 5×5 convolution in sequence. The 1×1 convolution achieves channel information fusion and dimensionality reduction, while the 5×5 convolution expands the local receptive field and extracts detailed target features. The outputs of the two branches are fused through Hadamard product and then adjusted to the target dimension by 1×1 convolution, outputting local enhanced features to improve the localization accuracy of tiny targets in complex backgrounds. The structural design of the TTFM in this embodiment is as follows: Figure 4 As shown, the model includes two parallel branches and a fusion operation: the input features are first transformed in dimension through feature reshaping and a fully connected layer. Then, a parallel branch uses a shifted window multi-head self-attention module (SW-MSA) to capture local contextual information. SW-MSA models local long-range dependencies through self-attention mechanisms within the window and expands the effective receptive field using a window shifting strategy, enhancing the model's ability to locate small targets and providing a clear local view for small target detection. Afterwards, a multilayer perceptron (MLP) is used to complete feature transformation and nonlinear mapping. In the other parallel branch, the convolutional branch, convolutional operations further enhance the extraction of local detail features: first, a 1×1 convolution is used to fuse and compress channel-dimensional information, and then a 5×5 convolution is used to expand the local receptive field, capturing a wider range of detail information, especially helping to accurately locate the boundaries of small targets in complex backgrounds. The output features of the two branches are added and fused. Finally, the fused features are reshaped in dimension through feature reshaping and a 1×1 convolution to adapt to the processing needs of subsequent network layers, forming the final output of the module.

[0019] The TTFM described above has significant adaptability advantages for small target detection in the airspace. Addressing the issue that small targets in the airspace occupy a very small percentage of pixels in images, relying solely on deep features from the backbone network can easily lead to the loss of crucial details, the additional small target branch retains rich low-level semantic information, perfectly meeting the feature extraction requirements for small targets in the airspace. The 7×7 shifted window size is adapted to the pixel scale of small targets in the airspace, and the window shifting strategy can effectively expand the local receptive field, accurately model the relationship between small targets in the airspace and their surrounding local context, and enhance the localization capability of small targets. The convolutional dual-branch can perform refined extraction of subtle texture and shape features of small targets in the airspace, adapting to the detailed features of small targets in the airspace such as drone fuselages and bird wings. The Hadamard product fusion of the dual-branch features achieves complementary enhancement of local context information and detailed features, significantly improving the localization accuracy of small targets in complex airspace backgrounds and solving the problems of ambiguous localization and high false negative rate in small target detection in the airspace.

[0020] S4. Implementation of Dynamic Feature Aggregation Module (DFAM): DFAM is set between the neck and the detection head. The specific parameters of the four feature extraction branches are as follows: First branch: 5×1 convolution, output channel number C2, activation function is ReLU; Second branch: 1×5 convolution, dilation rate Rate=2, output channel number C2, activation function is ReLU; Third branch: 5×5 dilated convolution, dilation rate Rate=2, output channel number C2, activation function is ReLU; Fourth branch: residual connection, directly passing the original input features; After concatenating the output features of the first, second and third branches, the number of channels is adjusted to C2 through 1×1 convolution, and then weighted and fused with the original features of the fourth branch through a learnable weight α. The initial value of α is set to 0.5 and is adaptively updated during training. The final output is a comprehensive feature that takes into account both local details and global context.

[0021] The structural design of the DFAM in this embodiment is as follows: Figure 5As shown, the input feature map is divided into two main paths. One main path is the basic feature branch: the input feature map first undergoes a 1×1 convolution to complete channel transformation, and then is multiplied by a weight 1-α as a global feature contribution. The other main path is the core of multi-scale feature extraction and fusion, and the process is as follows: First, after the input feature map is reduced in dimensionality by a 1×1 convolution, it is divided into two sub-paths. One sub-path directly uses a 1×1 convolution, and the other sub-path uses a 5×1 convolution with a stride of 2 to extract vertical features. The outputs of the two sub-paths are multiplied element-wise. Then, it is divided into two sub-paths again: one sub-path uses a 1×1 convolution, and the other sub-path uses a 1×5 convolution with a stride of 2 to extract horizontal features. The outputs of these two sub-paths are multiplied element-wise again. Subsequently, this result is concatenated with the outputs of the other two sub-paths, which are the convolution branch after the element-wise multiplication of the first part and the 5×5 convolution branch with a stride of 2, respectively. The concatenated feature is then processed by a 4×C2 channel 1×1 convolution. Convolutional compression and fusion are performed, followed by multiplication by a weight α. Finally, the feature maps of the two main paths are added element by element to obtain the output. α is adaptively updated during training, and the final output is a comprehensive feature that takes into account both local details and global context.

[0022] The DFAM described above has significant adaptability advantages for small airspace target detection. It is suitable for dense small targets such as drone swarms and bird flocks that often exist in airspace scenarios, and these targets have different flight attitudes and angles. Its multi-branch heterogeneous convolution design can specifically capture feature information of airspace targets in different directions. The 5×1 and 1×5 convolutions focus on the vertical and horizontal directions, respectively, adapting to the morphological features of airspace targets such as drone fuselages / wings and bird wings. The 5×5 dilated convolution expands the receptive field without losing details, and can aggregate pixel information of dense small airspace targets across a large range to extract global context features. Residual connections can avoid the degradation problem of small airspace target features in multi-layer fusion, ensuring the effective preservation of original features and realizing dynamic collaborative fusion of local details and global context. It optimizes feature representation in dense small airspace target scenarios and solves the problems of feature confusion and low discriminability in dense small airspace target detection.

[0023] S5. Dataset Construction and Preprocessing: VisDrone2019-DET Dataset: 6471 images for training, 548 images for validation, and 1610 images for testing, containing ten categories of targets including pedestrians, vehicles, and drones. Smaller target samples were selected for training. ASODD Dataset: 2863 bird and kite samples were selected from the COCO dataset, 3127 drone samples were extracted from the drone-detection-rchy7_dataset, and 2251 drone samples were collected independently, totaling 8241 images. Target categories and bounding boxes were uniformly labeled, and the images were divided into training, validation, and testing sets in a 7:2:1 ratio. All images were preprocessed: resized to 640×640 pixels, and data augmentation techniques such as random flipping, brightness adjustment, and Gaussian blur were used to improve the model's generalization ability.

[0024] S6. Model Training and Detection: Training Environment: PyTorch 1.18 framework, CUDA 12.4, dual NVIDIA GeForce RTX 3080 24GB cards; Training Parameters: Initial learning rate 0.01, cosine annealing learning rate scheduling strategy, momentum 0.937, weight decay 0.0005, batch size=8, maximum training epoch=200, joint optimization using cross-entropy loss function and GIoU loss function; Detection Process: The image to be detected is input into the trained GLFE-TDN network. After backbone network feature extraction, CFFM global enhancement, TTFM local enhancement, and DFAM feature fusion, the target detection results are output by four detection heads. Non-maximum suppression (NMS) is used to remove redundant detection boxes. Finally, the target category, bounding box, and confidence score with a confidence score greater than 0.5 are output.

[0025] In this embodiment, refer to Figure 2In the GLFE-TDN model, the image input is first fed into the backbone network. This network sequentially passes through two convolutional layers, then connects to the C3K2 module, another convolutional layer, and another C3K2 module. Next, it connects to the Context Feature Builder (CFFM) module, followed by another convolutional layer, then the A2C2F module, the CFFM module, another convolutional layer, and finally the A2C2F module. The features output from the backbone are fed into the neck network, which is built based on a bidirectional feature pyramid network. The neck network performs feature transformation through convolution and upsampling, and then performs multi-scale feature stitching. The neck network also contains A2C2F modules to further process the stitched features, achieving deep multi-scale feature fusion. Simultaneously, an additional micro-target branch is derived from the connection node between the backbone and the neck. This branch is specifically connected to the TTFM (Micro-Target Region Focusing Module) to enhance local features of micro-targets. The processed features are then stitched back to the corresponding feature layer in the neck, participating in the overall feature fusion of the neck. The Neck precisely feeds the fused multi-scale features to four detection branches: Large, Medium, Small, and Tiny. Each branch has a separate DFAM (Dynamic Feature Aggregation) module between the Neck output and the Detect module in the head, dynamically weighting and fusing the features from the neck with local details and global context. Finally, the DFAM-processed features from the four detection branches are input to their respective Detect modules, which output the target class probability, bounding box coordinates, and confidence score, forming a complete model feature extraction, enhancement, fusion, and detection chain.

[0026] In this embodiment, the global-local feature collaborative enhancement method for small target detection in the airspace has been validated for practical value in core scenarios such as anti-drone intrusion and airspace security monitoring: In low-altitude monitoring of airport runways, it can accurately identify distant flocks of birds and micro-drones, adapting to complex background environments; in military base no-fly zone early warning scenarios, it can detect small reconnaissance drones in complex environments such as low light at night and smoke interference, demonstrating outstanding anti-interference capabilities; in urban low-altitude control, it can accurately distinguish between drones, birds, and kites, adapting to diverse lighting and weather conditions. This method constructs a "global perception-local enhancement-dynamic fusion" link through CFFM, TTFM, and DFAM modules, and combines joint training with VisDrone2019-DET and the self-built ASODD dataset, overcoming the pain point of insufficient global-local feature collaboration in traditional algorithms, balancing real-time performance and generalization ability, and meeting the efficient detection needs of different scenarios such as sensitive areas and urban airspace.

[0027] The method in this embodiment was comprehensively compared and verified with current mainstream spatial target detection algorithms such as TPH-YOLO, Dynamic R-CNN, FFCA-YOLO, and YOLOv8m. The performance of the core detection indicators of each algorithm is shown in Table 1. The results show that the GLFE-TDN of this invention is superior to existing detection algorithms in terms of precision, mAP50, mAP50:95, and small target mAP, and also maintains excellent recall. Overall, it is significantly better than the comparison algorithms. The model demonstrates its stability in detecting airspace targets under different IoU thresholds. As a core indicator for small airspace target detection, the average accuracy of small targets in this invention is significantly improved compared to all mainstream algorithms, representing a significant breakthrough in small target detection accuracy. The performance comparison results fully demonstrate that this method effectively overcomes the technical bottleneck of insufficient global-local feature collaboration in traditional algorithms in small airspace target detection scenarios. Whether it is the detection accuracy of a single small airspace target or the comprehensive detection capability of small targets in complex airspace environments, it is comprehensively superior to existing mainstream detection algorithms. It has stronger robustness and adaptability in the classification and localization of small airspace targets such as drones, birds, and kites, and can better adapt to the detection needs of various practical application scenarios such as anti-drone intrusion and airspace security monitoring.

[0028] Table 1 shows the performance evaluation of different target detection methods; The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A spatial small target detection method with global-local feature co-enhancement, characterized in that, The method includes the following steps: S1. Constructing the overall network architecture: The backbone network processes the input image through deep feature extraction and generates a high-level feature representation; the neck uses a PAN bidirectional feature pyramid network for feature fusion; the detection head is set with four detection branches, corresponding to the detection of large targets, medium targets, small targets and tiny targets respectively; S2. Embed the Context Feature Aggregation Module (CFFM) in the backbone network: extract features in parallel through multi-scale convolution, and combine global average pooling (GAP) and global max pooling (GMP) to integrate channel and spatial context information to construct a global feature-dense representation. S3. An additional micro-target branch is introduced between the backbone network and the neck, and a micro-target region focusing module TTFM is embedded: local context information and detailed features are extracted through multi-head self-attention branches and convolutional branches. S4. Add a dynamic feature aggregation module DFAM in front of the detection head: adopt a multi-branch heterogeneous convolution design, dynamically weighted and fused local details and global context features of different receptive fields, and optimize feature representation in dense target scenes; S5. Constructing a training dataset: Combining the publicly available VisDrone2019-DET dataset with the self-built ASODD airspace small target dataset, the ASODD dataset contains labeled images of three types of targets: birds, kites, and drones. S6. Model Training and Detection: The model is trained using the stochastic gradient descent (SGD) optimizer. After feature processing by each module, the detection head outputs the target category, bounding box, and confidence score to complete the detection of small targets in the spatial domain.

2. The spatial small target detection method with global-local feature co-enhancement as described in claim 1, characterized in that, In S2, the CFFM includes a multi-scale parallel convolution branch and an attention enhancement branch: the multi-scale parallel convolution branch extracts multi-scale features, namely, it uses a 3×3 standard convolution, a 5×5 standard convolution, and a large receptive field convolution branch composed of a cascaded 1×7 depthwise separable convolution and a 7×1 depthwise separable convolution, which process the input features in parallel; at the same time, a direct cross-layer connection from the input to the multi-scale fused feature is designed to share information between different layers; at the end of the multi-scale parallel convolution branch, a 1×1 convolution is used to perform channel fusion and dimensionality unification on the parallel extracted multi-scale features; In the attention enhancement branch, global channel features are first extracted using both Global Average Pooling (GAP) and Global Max Pooling (GMP). The pooling results are then summed and normalized using Softmax to generate channel attention weights. This branch contains two sub-paths: the first sub-path performs matrix multiplication between the channel attention weights and the multi-scale enhanced features to generate global spatial modulation features; the second sub-path first reduces the dimensionality of the channel attention weights using a 1×1 convolution, then generates spatial attention weights through sigmoid activation, performs element-wise multiplication (Hadamard product) with the original features, and finally completes the spatial modulation feature transformation with another 1×1 convolution. Finally, the outputs of the two sub-paths are summed and residually concatenated with the original input features of the module to form the final output of the module.

3. A spatial small target detection method with global-local feature co-enhancement as described in claim 1 or 2, characterized in that, In S3, within the TTFM, the input features first undergo dimensionality transformation through feature reshaping and a fully connected layer. Then, a parallel branch employs a multi-head self-attention (SW-MSA) module with a shifted window to capture local contextual information. SW-MSA models local long-range dependencies through a self-attention mechanism within the window and expands the effective receptive field using a window shifting strategy, enhancing the model's ability to locate small targets and providing a clear local view for small target detection. Afterward, a multilayer perceptron (MLP) is used to complete feature transformation and nonlinear mapping. In another parallel branch, the convolutional branch, convolutional operations further enhance local detail feature extraction: first, a 1×1 convolution is used for channel-dimensional information fusion and compression; then, a 5×5 convolution expands the local receptive field, capturing a wider range of detail information, especially helpful in accurately locating the boundaries of small targets in complex backgrounds. The output features of the two branches are added and fused. Finally, the fused features undergo feature reshaping and 1×1 convolution for dimensionality adjustment to adapt to the processing requirements of subsequent network layers, forming the module's final output.

4. A spatial small target detection method with global-local feature co-enhancement as described in claim 1 or 2, characterized in that, In S4, in the DFAM, the input feature map is divided into two main paths, one of which is the basic feature branch: The input feature map first undergoes a 1×1 convolution to transform the channels, then is multiplied by a weight of 1-α as a global feature contribution. The other main path is the core of multi-scale feature extraction and fusion, with the following process: First, the input feature map undergoes dimensionality reduction via a 1×1 convolution, then splits into two sub-paths. One sub-path directly uses a 1×1 convolution, while the other uses a 5×1 convolution with a stride of 2 to extract vertical features. The outputs of these two sub-paths are multiplied element-wise. Then, it is again split into two sub-paths: one uses a 1×1 convolution, and the other uses a 1×5 convolution with a stride of 2 to extract horizontal features. The outputs of these two sub-paths are multiplied element-wise again. Subsequently, this result is concatenated with the outputs of the other two sub-paths, which are respectively the convolution branch after the element-wise multiplication of the first part and the 5×5 convolution branch with a stride of 2. The concatenated feature map is then processed by a 4×C2 channel 1×1 convolution. The convolutional compression and fusion are then multiplied by a weight α. Finally, the feature maps of the two main paths are added element by element to obtain the output. α is adaptively updated during training, and the final output is a comprehensive feature that takes into account both local details and global context.

5. A spatial small target detection method with global-local feature co-enhancement as described in claim 1 or 2, characterized in that, In S5, the ASODD dataset is constructed as follows: N1 bird and kite samples are selected from the COCO dataset, N2 drone samples are extracted from the drone-detection-rchy7_dataset, and N3 drone samples are collected autonomously, for a total of N1+N2+N3 images. The target categories and bounding boxes are uniformly labeled.

6. The spatial small target detection method with global-local feature co-enhancement as described in claim 5, characterized in that, In step S5, the data is divided into training set, validation set, and test set according to a set ratio; all images are preprocessed: resized to 640×640 pixels, and data augmentation is used to improve the model's generalization ability.

7. The spatial small target detection method with global-local feature co-enhancement as described in claim 6, characterized in that, In S5, the data augmentation methods include random flipping, brightness adjustment, and Gaussian blur.

8. A spatial small target detection method with global-local feature co-enhancement as described in claim 1 or 2, characterized in that, In S6, the training parameters are set as follows: PyTorch 1.18 framework, CUDA 12.4, dual NVIDIA GeForce RTX 3080 24GB cards are used; the initial learning rate is set, the cosine annealing learning rate scheduling strategy is adopted, and the momentum, weight decay, batch size and maximum training epoch are set. The cross-entropy loss function and GIoU loss function are jointly optimized.

9. The spatial small target detection method with global-local feature co-enhancement as described in claim 8, characterized in that, In S6, the detection process is as follows: the image to be detected is input into the trained GLFE-TDN network, and after backbone network feature extraction, CFFM global enhancement, TTFM local enhancement, and DFAM feature fusion, the target detection results are output by four detection heads. Non-maximum suppression is used to remove redundant detection boxes, and finally the target category, bounding box and confidence score with a confidence score greater than the threshold are output.

10. The spatial small target detection method with global-local feature co-enhancement as described in claim 8, characterized in that, In S6, the initial learning rate is 0.01, the momentum is 0.937, the weight decay is 0.0005, the batch size is 8, the maximum training epoch is 200, and the threshold is 0.5.

Images